EEG-Ferro

Code for the forthcoming paper:

Personalized Spiking Neural Networks with Ferroelectric Synapses for EEG Signal Processing

Nikhil Garg, Anxiong Song, Niklas Plessnig, Nathan Savoia, and Laura Bégon-Lours

Integrated Systems Laboratory, ETH Zürich

Paper link: https://doi.org/10.1063/5.0319912.

Overview

We train a spiking neural network (SNN) to classify left-hand vs. right-hand motor imagery from EEG signals, targeting deployment on neuromorphic hardware with ferroelectric FET (FeFET) synaptic devices. The pipeline consists of four steps:

1. Floating-point pretraining — Train the SNN on pooled multi-subject data using standard backpropagation.
2. On-device training — Retrain from scratch using a device-aware optimizer that models FeFET conductance updates via a beta-distribution kernel.
3. On-device finetuning — Load pretrained weights, quantise and add device noise, then finetune with the device model.
4. Subject-specific transfer learning — Adapt the model to individual subjects using only their own data (online, batch_size=1).

Repository Layout

dataset/              EEGMMIDB preprocessing, CV splits, and motor imagery dataset loaders
src/                  Reusable training, model, and device code
  models/             SNN architecture
  training/           Training, on-device learning, transfer learning, re-tuning
  devices/            Quantization
scripts/              Thin command-line entry points
configs/              Experiment parameters
results/              Figures, postprocessed CSVs, logs, checkpoints
notebooks/            Analysis and plotting notebooks

Paper-to-Code Map

Paper section	Repository location
Motor imagery dataset	dataset/
SNN architecture	src/models/
Floating point baseline	src/training/train_floating_point.py
On-device learning	src/training/train_on_device.py
Subject-specific transfer learning	src/training/SSTL.py
Quantization/noise	src/devices/quantization.py
Re-tuning	src/training/retuning.py
Figures/results	results/ and notebooks/

Core Files

File	Role
src/models/snn.py	CUBA-LIF spiking CNN architecture
src/training/train_floating_point.py	Floating-point baseline training
src/training/train_on_device.py	Device-aware on-device learning
src/training/SSTL.py	Subject-specific transfer learning
src/training/eval_pretrain.py	Evaluating per-subject accuracy of pretrained model
src/training/retuning.py	Low-overhead device-aware re-tuning
src/devices/quantization.py	Quantization and programming-noise utilities
configs/config.py	Experiment arguments and defaults

Results and Notebooks

Generated SVG figures are under results/figures/. Postprocessed CSVs are under results/postprocessing/, and raw experiment outputs/checkpoints are under snn_experiment_results.

Plotting and analysis notebooks are grouped under notebooks/figures/, Notebook for analysing MAC count is under notebooks/MAC/

SNN Architecture

The network processes 2D spatial EEG maps (10×11 grid, 160 timesteps) through:

Layer	Type	Output Shape	Description
Conv1	Conv2d + CUBA-LIF	64 × 8 × 9	3×3 convolution
Conv2	Conv2d + CUBA-LIF	128 × 6 × 7	3×3 convolution
Pool	AvgPool2d	128 × 3 × 3	2×2 average pooling
Conv3	Conv2d + CUBA-LIF	256 × 1 × 1	3×3 convolution
TC1	Temporal Conv + LIF	256	Kernel=3, aggregates 3 timesteps
R1	Recurrent LIF	256	Adaptive threshold
FC1	Linear + LIF	256	With dropout
FC2	Linear (readout)	2	Weighted sum over timesteps

All hidden layers use CUBA-LIF (Current-Based Leaky Integrate-and-Fire) neurons with learnable per-neuron current and voltage decay constants and rectangular surrogate gradients for backpropagation.

Key Parameters

Neuron Parameters (Eq. 2–5)

Parameter	Symbol	Default	Description
CDECAY	β	0.1	Current decay constant (Eq. 2)
VDECAY	γ	0.1	Voltage decay constant (Eq. 3)
VTH	v_th	0.1	Spike threshold (Eq. 4)
GRAD_WIN	g	0.3	Surrogate gradient window width (Eq. 5)
TH_AMP	—	0.01	Adaptive threshold increment
TH_DECAY	—	0.1	Adaptive threshold decay
BASE_TH	—	0.01	Baseline threshold

Device Model Parameters (config.py, Eq. 14–15)

Argument	Symbol	Default	Description
--dut	—	dut2	Device-under-test (dut1–dut5)
--thr	ε	0.025	Weight update threshold as fraction of range (Eq. 15)
--thr_asym	ε_asym	1.0	LTP/LTD threshold asymmetry (Eq. 15)
--x	1/η	4.0	Noise scaling; η = 1/x is the RSD (Sec. 3.4)
--q	q	0	Quantization levels (0 = full precision)
--A_scale	—	1.0	Global conductance update scaling
--A_asym	—	1.0	LTP/LTD amplitude asymmetry
--rsd_c2c	σ_c2c	0.0	Cycle-to-cycle variation (RSD on A)

Training Parameters

Argument	Default	Description
--lr	1e-4	Pretraining learning rate
--finetune_lr	2e-4	Finetuning learning rate
--pretrain_epochs	20	Number of pretraining epochs
--retuning_epochs	4	Number of re-tuning epochs
--pretraining_batch_size	64	Batch size for pretraining
--retuning_batch_size	64	Batch size for re-tuning
--finetuning_batch_size	1	Batch size for SSTL
--finetune_epochs	5	Number of SSTL epochs
--seed	5	Random seed

Setup

1. Install dependencies

pip install -r requirements.txt

2. Download the dataset

Download the PhysioNet EEG Motor Movement/Imagery Dataset into a directory named "data":

mkdir -p eegmmidb/physionet
# Download using your preferred method, e.g.:
# wget -r -np https://physionet.org/files/eegmmidb/1.0.0/

3. Preprocess the data

python dataset/utility.py

This preprocesses and saves normalised, sliced EEG data as pickle files under eegmmidb_slice_norm/.

Reproducing Paper Results

Run the steps in order:

# Floating point pretraining
python src/training/train_floating_point.py

# On-device training with FeFET model
bash scripts/run_train_on_device.sh

# Finetune pretrained model after simulated weight transfer
bash scripts/run_retuning.sh

# Subject-specific transfer learning (SSTL)
bash scripts/run_SSTL.sh

# Evaluating per-subject accuracy of pretrained model on held-out subject (for calculating delta accuracy after SSTL)
python src/training/eval_pretrain.py

Each script saves CSV log files and .pth model checkpoints to the results/ directory.

Dataset

• Name: PhysioNet EEG Motor Movement/Imagery Dataset (EEGMMIDB)
• Subjects: 109 (6 excluded due to data quality: 88, 89, 92, 100, 104, 106)
• Task: Left-hand vs. right-hand motor imagery
• Sampling rate: 160 Hz
• EEG channels: 64, mapped to a 10×11 spatial grid
• Epoch length: 160 samples (1 second)

Citation

@article{garg2026Personalized,
    author = {Garg, Nikhil and Song, Anxiong and Plessnig, Niklas and Savoia, Nathan and Bégon-Lours, Laura},
    title = {Personalized spiking neural networks with ferroelectric synapses for EEG signal processing},
    journal = {APL Machine Learning},
    volume = {4},
    number = {2},
    pages = {026110},
    year = {2026},
    month = {05},
    issn = {2770-9019},
    doi = {10.1063/5.0319912},
    url = {https://doi.org/10.1063/5.0319912},
    eprint = {https://pubs.aip.org/aip/aml/article-pdf/doi/10.1063/5.0319912/21009355/026110_1_5.0319912.pdf},
}